Vinod Kumar Reddy Gandra

Built a bot which gives pointed answers for natural language questions from a corpus of budget article data. Used a pipeline of TfIdf-based retrieval and glove vector-based recurrent neural network system to answer the queries. Optimized the pipeline to decrease the latency and increase the throughput to fit the production requirements of MoneyControl. Achieved a throughput of 100 requests/second with a latency of one second. Improved the performance of pipeline by incorporating Solr instead of the tfidf-based retriever and text summarizer to better show the results. Achieved a top accuracy of 75% on test data.

Modeled user intent from text chat in a bot/human conversation as part of that diagnostic system. Benchmarked several NLU systems like DeepPavlov, ParlAI, RasaNLU, and Stanford NLP on a custom chat corpus. Built NER models to extract relevant entities in a human and bot chat and built sentiment models to figure out the user sentiment with reference to the entities.

Implemented a YOLO-based people counting system to detect crowds in the view of security cameras.

Implemented detecting masked faces in the view of security cameras. Trained and deployed a face attention network on the custom dataset to detect the masked faces.

Built models on order level transactional data to predict whether an order ends up asRTO/Return. Used several ML techniques to generate user and item level features, historic, demographic and time series features to improve model accuracy. Modeled user behavior from user click data to improve the RTO prediction model by 20%. Built an RNN based model to parse a delivery address into individual components and use the subsequent features to improve the model of the RTO prediction model.